# One State's Power Permits Against the 2028 Compute Pipeline — Listening to MacroMicro EP. 221

Source: Realpha Blog (blog.getrealpha.com)
Original article and charts: https://blog.getrealpha.com/en/blog/macromicro-2026-11-08-ai-after-meeting-ep-221/

> MacroMicro's After Meeting EP.221 (November 2026) takes the US midterms — Democrats winning both chambers, Texas flipping — and translates the result into a question about data centre power permits. Educational listening notes on how to read a moratorium and why free cash flow turned negative — no price targets, no stock recommendations.

Published: 2026-11-08
Locale: en
Tags: midterm elections, data centres, AI infrastructure, free cash flow, divided government

TL;DR: The 2027 compute is already under construction, and existing moratoria clip only about 2.3GW of it — some 6%. The variable is the 62GW slated for 2028, roughly 30GW of which waits on states legislating first and permitting after.


![A Texas suburban street at dusk, a sprinkler turning on dry grass in the foreground, a chain-link fence and young trees mid-frame, and a half-finished data centre campus with transmission pylons at the far end of the road](/covers/macromicro-2026-11-08-ai-after-meeting-ep-221-cover.png)

> In the compound republic of America, the power surrendered by the people is first divided between two distinct governments... Hence a double security arises to the rights of the people. ... Ambition must be made to counteract ambition.
>
> — James Madison, *The Federalist* No. 51 (1788)

MacroMicro's After Meeting EP. 221, published on 8 November 2026 and recorded on the afternoon of 5 November, has host Roger and research manager Dylan on the US midterms: the Democrats took both chambers, with 228 House seats already confirmed and 20 of the 35 Senate seats up for election, and the biggest surprise was Texas flipping both its governorship and its Senate seat. On the day the first results landed the Dow rose 3% while the Philadelphia Semiconductor Index fell more than 3% to its lowest since April, because what the market was pricing was whether data centre power permits would now be held up by new state governments. Dylan's reading is that the data centres due to finish in 2027 are already being built, with existing moratoria affecting about 2.3GW or 6%, and that the variable is the 62GW coming in 2028, roughly 30GW of which needs state approval — a conclusion that holds as long as the five big cloud providers keep capital spending on its path toward $1–1.2 trillion, money local governments find hard to turn down.

## An episode taped while the count was still running

This one was recorded at 1:30pm Taiwan time on 5 November, with a few districts still counting. Roger opens on the sharp drop in Taiwan's margin maintenance ratio, then mentions patrolling the forums and finding people comparing which park is more comfortable to live in — the way people stuck in losing positions joke about it here — before adding that if you size your positions properly you do eventually get back to the hotel.

US equities split in two that day. The Dow rose 3% and the S&P 500 1.6%, with healthcare, hospitals and renewables up, while anything tied to data centre construction weakened: CoreWeave closed down 10%, Tesla and SpaceX, both with large Texas footprints, fell 3–4%, and the Philadelphia Semiconductor Index fell more than 3% to a new low since April. Elon Musk posted the all-blue Texas results map on X with a line about data centres having to go to space after all.

![On a single day the market splits from one shared starting point into two lines, one rising for the Dow and one falling for the Philadelphia Semiconductor Index, down to its lowest level since April.](/figures/election-day-market-split-en.svg)

## Both chambers flipping was expected; Texas was not

Dylan says both chambers turning blue sat inside expectations — polls and event markets like Polymarket had signalled it. Texas was the surprise: Republicans still held a slight edge going in, and lost the governorship and the Senate seat together. He compresses the election into two words, divided government and Texas.

Three numbers sit underneath divided government. A Senate majority is 51 seats, but there is also 60: a senator who keeps talking stops a bill reaching a vote, and ending that kind of extended speech takes 60 votes, which neither side can reach. The third number is two-thirds — the president can veto a bill passed by Congress, overriding a veto takes two-thirds, and the Democrats are short of that too. Trump's first term from 2019 to 2021 was already a divided government, and he used the veto ten times.

![Three threshold lines rise from low to high; the Democratic seat bar clears the lowest majority line but stops below the two lines above it.](/figures/senate-three-thresholds-en.svg)

So will tariffs and foreign policy turn? Dylan's answer is no: the 60 and the two-thirds both constrain the Democrats, tariffs and foreign policy are executive territory to begin with, and Trump prefers governing by executive order. What the new Congress does change is three things involving money — the continuing resolution expiring on 11 December (the government shutdown line), the debt ceiling that gets reached somewhere between February and July 2027, and control of budget allocation. On the first two the Democrats have reason to stall until the new Congress is seated in January.

## What people object to is the electricity bill, the water, and an NDA

The part of this episode I did not expect was why Americans oppose data centres. Dylan lists three reasons: the electricity bill (energy prices were already rising this year, data centres eat power, and residents readily pin a higher bill on them), water (states short of it, like Arizona and Texas, are sensitive), and one I had not thought about — data centres are usually built under non-disclosure agreements, so when a local council chases the water and power numbers it cannot get them, and what voters see is a black box.

In this cycle's exit polling about 76% of respondents leaned against data centres, and that 76% is cross-party. Dylan mentions the campaign slogan Not in my backyard — your data centre is fine, just not next to my house — and the Texas version, Texas not for sale, meaning my water and my power are not yours to sell off. He describes the election as a referendum on trust in AI infrastructure.

Regulation works on two levels. Dylan says the federal one does not worry him, with the 60 votes and the veto both in place; what worries him is the states, and Texas most of all. Texas alone holds about 104GW of power capacity across operating, under construction and planned — and seven-tenths of that is still in planning, roughly three to four times what is operating and building today. The incoming Democratic governor has already said pause first, legislate after.

![Texas power capacity bar: 30% is operating or under construction, while 70% is planned and behind the new governor’s permit gate.](/figures/texas-capacity-stack-en.svg)

The pressure was building before the vote. New York and California had already moved to tighter review or outright moratoria, and in September Texas's then-Republican governor asked for a halt on new permits until a grid audit was complete. The election pushed the trend one step further: across the ten key states Dylan compiled, every governor-elect has signalled a pause or tighter regulation, and the harder-line ones cluster in the big data centre states. Florida, which the Republicans held, has begun asking operators to carry costs and is removing some tax breaks, though Florida's capacity is only 0.9GW. Beyond governors there are the local ballot measures: of the several dozen he went through, exactly one leaned toward building.

## 2027 loses 6%; the variable is 2028

This is the most useful distinction in the episode. Existing moratoria generally apply to new applications only and are not retroactive, so the data centres due to finish in 2027 broke ground long ago. Dylan cites SemiAnalysis: the New York and other bans add up to roughly 2.3GW, about 6% of 2027.

![A gate stands on a timeline: the arrow for projects already under construction passes through from the left and reaches completion, while the arrow for new applications stops against the gate's right side.](/figures/moratorium-not-retroactive-en.svg)

2028 looks different. The estimate is 62GW of additions, most of it not yet started and needing state approval to move, with Texas accounting for three-tenths and governors in Georgia and New Mexico also talking about legislating or pausing first. Dylan's figure is about 30GW exposed — capacity that may get re-litigated or slowed, which is not the same as 30GW halted.

![Two bars side by side for two years of new capacity: the shorter one has only a small slice at the top marked, while nearly half of the taller one is marked.](/figures/2027-vs-2028-exposure-en.svg)

The other change is that power becomes the operator's own problem. It used to come from the grid; now cloud providers share grid-upgrade costs, carry water and power bills, and face fewer tax breaks and longer build times. Dylan's way of putting it is that the entry ticket got more expensive. What this hits is the speed and cost of building data centres, not AI demand.

![Two stacks of blocks, before and now: the one for now piles several more cost layers on top of the original grid power, making it clearly taller.](/figures/entry-ticket-stack-en.svg)

## Five counterpoints: it may just get built in a different state

Dylan then gives five arguments against his own setup, and I think this stretch is worth more than anything before it.

First, restricting data centres is a cross-party consensus that started before the election — Texas's Republican governor did it — so the incremental difference from a Democratic win may be smaller than the market assumed. Second, what Democrats are asking for is review; nobody is asking for a blanket halt, and the condition written into the pauses is that new permits can be considered once legislation is done. Third, new capacity has other states to flow to: by the Smart Electric Power Alliance's count, 37 states and about 104 large-load programmes are planning or reviewing dedicated tariffs, and states with clear standards will pull projects in. Fourth, at federal level both parties share the worry about China gaining an AI edge, and everything from chip export controls to restrictions on Chinese firms renting cloud compute has had bipartisan sponsors. Fifth, which he calls the key one: campaign language runs hot, and candidates of either party will follow the public mood — yet the five big cloud providers' capital spending is heading toward $1–1.2 trillion, and local governments find an investment that size hard to refuse.

## The column falling this quarter is free cash flow

The second half turns to earnings. More than seven-tenths of the S&P 500 has reported, with overall profit growth near 30%, while the Magnificent Seven's estimated earnings per share got revised down and now lag the other 493 — the first time, Roger notes, since the fourth quarter of 2022.

He puts the ugly EPS down to one-offs: Alphabet booked a $94 billion gain on its SpaceX investment last quarter, and Meta booked $9.9 billion of litigation and settlement costs. What to watch is cash flow. Capital spending may reach $800 billion this year against $710 billion of full-year operating cash flow, so free cash flow has already turned negative, roughly −$35 billion in the third quarter. Three more pressures push the same way: depreciation climbing, this year's debt issuance starting to accumulate interest next year, and off-balance-sheet commitments now estimated above $4 trillion.

![The bar for cash coming in falls short of the bar for capital spending, and the difference opens a gap below the baseline.](/figures/free-cash-flow-turns-negative-en.svg)

By the end of October, forward price-to-earnings multiples for Amazon, Alphabet and Oracle were down to 21, 17 and 13. Dylan's description is that these companies are turning from asset-light cloud platforms into capital-heavy, highly levered data centre operators, and the market is pricing the latter.

He also gives the path back. Capital spending growth peaks this third quarter at 117% and falls to about 45% in 2027, while cloud revenue growth at Google Cloud, Azure and AWS is still rising this quarter. Net the two and he puts free cash flow back in positive territory in 2028, provided AI demand holds up.

Anthropic's amended S-1 filed this week supplies a counter-intuitive example. Revenue is about 13 times the same period last year, yet the third quarter went to a loss, and Dylan's main reason is $5 billion of short-term compute spending: long-term contracted supply cannot keep up with demand, so in the short run it has to buy expensive compute on the spot. The loss came from not being able to buy cheap compute. The same filing carries compute and infrastructure commitments close to $900 billion, and if revenue fails to keep pace that becomes Anthropic's problem.

![A falling spending-pace curve and a rising revenue-pace curve meet on the right, with the crossing point labeled as the turn back to positive.](/figures/two-curves-crossing-en.svg)

## Fourteen midterms, a 100% one-year hit rate

Dylan closes with a backtest. Across the 14 midterms since 1970, the year leading in was flat for equities, and the S&P 500 was higher one year after the vote in every single case; narrow the sample to divided governments and the hit rate is still 100%. His explanation is that what companies fear is uncertainty, and an election removes it. Within divided government he also flags Treasury yields having room to fall, since two parties pulling against each other on fiscal policy makes large-scale expansion difficult.

The sample is 14, and the rate and valuation backdrop differed every time, so I read that 100% as a sense of direction rather than a probability.

## The election news is out — what should I buy and sell?

The first reaction to a headline like "Democrats sweep both chambers" is usually to go looking for winner and loser lists: Democrats do not favour expanding defence, so defence should be weak; they favour health insurance, so healthcare should be strong. This episode walks a finer route.

Dylan's method is to find the physical thing the election result can actually touch. Winning Congress touches the order of budget allocation. Winning the Texas governorship touches the issuing of permits for new data centres, and what holds a permit up is power and water. Only from there does it connect to capacity, to the 62GW and the 30GW, to component lead times. Every link in that chain has someone signing, a clause, a number you can look up — while "the Democrats are anti-AI" has nobody signing anything.

![The top row is a chain of interlocking links, each with a signature mark below it; the bottom row holds a single isolated link with nothing to connect to.](/figures/chain-with-signatures-en.svg)

So with policy news I now ask two questions first: which physical thing does this change (a permit, a budget line, a pipeline, a tariff), and who signs it. The claim in this episode that answers neither is "the Democrats are anti-AI."

## Profits grew 30%, so why is my stock down?

Profit growth near 30% while the stock in your account falls is a punishing gap. The explanation here is that profit describes last quarter, whereas depreciation, interest and off-balance-sheet commitments all land their weight in later years; none of them enter this quarter's EPS, and all of them decide the cash that follows.

Two columns are worth tracking. One is free cash flow: turning negative is not itself a failure, but it needs a timetable for turning back and the conditions that timetable rests on. Dylan's version is 2028, and his condition is explicit — the spending curve (capex growth going from 117% to 45%) crossing the earning curve (cloud revenue growth still rising). Both are separately trackable, and the moment cloud revenue growth turns over the whole inference needs redoing.

The other is whose money it is. His approach to picking cloud providers is to take the ones already generating cash, since the ones leaning on outside financing carry more risk. The nice thing about this column is that you do not have to wait for a filing: look at whether the company has lately been issuing debt, selling shares, or collecting cash.

## Further reading

- MacroMicro, After Meeting EP. 221 (recorded 5 November 2026, published 8 November), hosted by Roger with research manager Dylan; the midterms and data centres run through the first half, earnings through the second
- MacroMicro's November 2026 monthly report, reading the four big cloud providers' results and AI capital spending
- MacroMicro's AI supply chain report, which separates long-lead from short-lead components — the basis for the segment on avoiding long-lead parts
- SemiAnalysis's estimate of the capacity affected by 2027 data centre bans
- The Smart Electric Power Alliance's count of large-load dedicated tariff programmes across the US
- James Madison, *The Federalist* No. 51 (1788, public domain), on the double division of power between federal and state

## One thing to take away

One idea: when something gets stopped, go read the conditions written for starting it again. If the conditions can be stated, it was postponed; if they cannot, that is when it was cancelled — and in a headline the two look identical. The whole inference in this episode grows out of that distinction: the moratoria cover new applications only and are not retroactive, so the 2027 volume proceeds, while the condition for resuming is legislation first and permits after, which is why the pressure lands on the 2028 capacity that has not broken ground.

Here is something I have tried, and it works outside investing: pick something in your own life that has been put on hold — the role your company froze, the thing a family member said we would talk about after the holidays, the test a doctor said to watch for now — and write three lines on paper. What has to happen before it resumes, who signs that, and the earliest date it could be. If all three come out, what you hold is a timetable. If one comes out blank, that blank is the question to ask today. I have done this a few times, and the line that comes up empty most often is the second one; once I have asked who signs, the third line usually answers itself.